TRACE manual - Center for Statistical Genetics

TRACE :
fasT and Robust Ancestry Coordinate Estimation
version 1.01
Chaolong Wang1
Department of Biostatistics
School of Public Health
Harvard University
April 9, 2015
The TRACE software2 is available at
http://www.sph.umich.edu/csg/chaolong/LASER/
1 Comments
2 This
on the TRACE software can be sent to [email protected].
software is licensed under the GNU General Public License, version 3.0.
Contents
1 Introduction
2
2 Getting started
2.1 Availability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Installing TRACE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Running TRACE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
3
3
3
3 Examples
3.1 Basic usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Parallel jobs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Internal reference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4
4
6
6
4 Input files
4.1 parameterfile ( .conf ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 studyfile ( .geno), genotypefile ( .geno), and sitefile ( .site) . . . . . . . . . . .
4.3 coordinatefile ( .coord ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
7
7
8
5 Usage options
5.1 Main parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Advanced parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Command line arguments . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
9
10
11
6 Output files
6.1 .log and terminal outputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 .RefPC.coord and .RefPC.var . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3 .ProPC.coord . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
13
13
14
7 Computational complexity
14
8 Version changes
8.1 Version 1.0 (May 19, 2014) . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2 Version 1.01 (Apr 6, 2015) . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
15
15
9 Acknowledgements
15
References
16
1
1
Introduction
TRACE is a software program that uses SNP genotypes to trace individual ancestry by comparing to a set of reference individuals. The basic idea is to construct a reference ancestry map
by applying principal components analysis (PCA) on genotypes of the reference individuals,
and then to place the study samples one by one into this reference ancestry map. With an
appropriate reference panel, the estimated coordinates of the study samples are informative
on their ancestry background. To place each sample, we use a two-step approach involving
PCA and projection Procrustes analysis (Gower and Dijksterhuis, 2004; Wang et al.,
2010). The TRACE program follows the same framework as the LASER program, which
we previously developed to estimate individual ancestry by directly analyzing shortgun reads
from next generation sequencing without calling genotypes (Wang et al., 2014). Different
from LASER, TRACE takes genotype data as the input, and thus can benefits studies when
sequence read data are not available.
TRACE can identify population structure in large cohorts. Compared to standard PCA,
TRACE has several advantages in terms of robustness and computational efficiency. First,
TRACE is robust to missing data in the study samples. Standard PCA requires high quality
genotypes with low missing data rate for the study samples. Commonly used PCA software
programs such as smartpca (Patterson et al., 2006) often impute missing data with mean
values at the corresponding loci of the genotype matrix. Consequently, samples that have
more missing data will shrink closer to the center of the PCA map. TRACE, in contrast, only
requires low missingness in the reference individuals and can appropriately handle missing
data in the study samples, which makes TRACE applicable to low quality data such as those
obtained from ancient DNA samples (Skoglund et al., 2012). Second, TRACE is robust to
the presence of close genetic relatedness between study samples. If PCA is applied direclty on
all study samples, closely related individuals often appear as outliers, and might distort the
overall PCA pattern. TRACE avoids this problem by analyzing each sample independently
with the reference panel, which only requires the reference individuals to be unrelated with
each other and with the study samples. For the same reason, TRACE is also robust to
uneven sampling scheme and existence of genetic outliers in the study samples, which are
known to have large impacts on PCA results (McVean, 2009; Lee et al., 2010). Last, the
computational time of TRACE increases linearly with the number of study samples, so that
TRACE can be much faster than standard PCA when the sample size is large.
Details of the method and examples illustrating the aforementioned advantages of TRACE
can be found in our paper entitled “ Improved ancestry estimation for both genotyping and
sequencing data using projection Procrustes analysis and genotype imputation” (Wang et al.,
2015).
2
2
Getting started
2.1
Availability
The TRACE program is distributed as part of the LASER software package (Wang et al.,
2014, 2015), which can be downloaded from the following webpage: http://www.sph.umich.
edu/csg/chaolong/LASER/. In the package, we provide a pre-compiled executable for TRACE
for Linux (64-bit) operation systems. This program is licensed under the GNU General Public
License, version 3.0. A copy of the license is included in the package or can be found at
http://www/gnu.org/licenses.
Source code written in C++ is provided in the package. The TRACE program uses two
external C++ libraries: the Armadillo Linear Algebra Library (http://arma.sourceforge.
net) (Sanderson, 2010) and the GNU Scientific Library (http://www.gnu.org/software/
gsl). The Armadillo library requires two additional libraries: LAPACK and BLAS. Therefore,
to compile from source code, you need to have these four libraries installed in your computer.
Please cite the following papers when you use TRACE :
1. Wang et. al. (2014) Ancestry estimation and control of population stratification in
sequence-based association studies. Nature Genetics, 46: 409-415.
2. Wang et al. (2015) Improved ancestry estimation for both genotyping and sequencing
data using projection Procrustes analysis and genotype imputation. AJHG, (accepted).
2.2
Installing TRACE
Open a terminal in the same directory as the .tar.gz file. Extract the file by typing
tar -xzvf LASER-2.02.tar.gz in the terminal. This will create a new directory called
LASER-2.02. This direcotry contains executables for both LASER (version 2.02) and TRACE
(version 1.01), and other resource files.
2.3
Running TRACE
Open a terminal and path to the directory that contains the executable TRACE. If you did
not rename the directory after extracting the .tar.gz file, the directory will be LASER-2.02.
Execute the program by typing ./trace -p parameterfile , in which -p is the command
line flag specifying the parameter file and parameterfile is the name of the parameter file.
If your parameterfile is not in the same directory, you must specify the whole path to the
file. If the parameterfile is not specified, TRACE will search in the current directory for a
parameterfile named “trace.conf”, and execute the program with parameter values specified in
“trace.conf”. If this file does not exist, an empty template parameterfile named “trace.conf”
will be created in the current directory. For more command line arguments, see Section 5.3.
3
3
Examples
This section provides example usage of the TRACE program based on data in the folder
named “example” (included in the download package). If you have questions when reading
this section, please refer to the next few sections for detailed information about the input files
(Section 4), usage options (Section 5), and output files (Section 6).
The “example” folder contains genotype data across 9,608 SNPs on chromosome 22 for 938
individuals from the Human Genome Diversity Panel (HGDP, Li et al., 2008). The HGDP
data are splited into two files named HGDP 700 chr22.geno and HGDP 238 chr22.geno, containing 700 and 238 individuals, respectively. The HGDP chr22.site file provides detailed
information of the 9,608 SNP markers. The HGDP 238 chr22.RefPC.coord file contains PCA
coordinates for the top 8 PCs based on genotypes in the HGDP 238 chr22.geno file. The folder
also includes a parameterfile named “example.conf”, which specifies parameters for running
TRACE on the example data. In the following examples, we will use the subset of 238 individuals to construct the reference PCA space and use TRACE to place the remaining 700
individuals into the reference PCA space.
3.1
Basic usage
After decompressing the download package, enter the folder that contains the executable
TRACE program. The following command will use parameter values provided in the example
parameterfile (shown at the end of this section).
./trace -p ./example/example.conf
The following command will change the number of PCs (DIM ) to 6 and the prefix of output file names (OUT PREFIX ) to “HGDP”, while the other parameters are defined by the
example.conf file.
./trace -p ./example/example.conf -o HGDP -k 6
Results from TRACE will be output to the current working directory.
The example parameterfile is similar to the one shown below. Each line specifies one
parameter, followed by the parameter value (or followed by a “#” character if the parameter
is undefined). Text after a “#” character in each line is treated as comment and will not be
read by the program.
# This is a parameter file for TRACE v1.01.
# The entire line after a ‘#’ will be ignored.
###----Main Parameters----###
STUDY_FILE
./example/HGDP_700_chr22.geno
4
# no default value
GENO_FILE
./example/HGDP_238_chr22.geno
# no default value
COORD_FILE
./example/HGDP_238_chr22.RefPC.coord
# no default value
OUT_PREFIX
test
# default "trace"
DIM
2
# default 2
DIM_HIGH
20
# default [20]
MIN_LOCI
100
# default 100
###----Advanced Parameters----###
ALPHA
# default 0.1
THRESHOLD
# default 0.000001
REF_SIZE
# default 200
FIRST_IND
# default 1
LAST_IND
# default [last sample in the STUDY_FILE]
TRIM_PROP
# default 0
MASK_PROP
# default 0
EXCLUDE_LIST
# no default value
PROCRUSTES_SCALE
# default 0 [include scaling as a free parameter]
###----Command line arguments----###
# -p
parameterfile (this file)
# -s
STUDY_FILE
# -g
GENO_FILE
# -c
COORD_FILE
# -o
OUT_PREFIX
# -k
DIM
# -K
DIM_HIGH
# -l
MIN_LOCI
# -a
ALPHA
# -t
THRESHOLD
# -N
REF_SIZE
# -x
FIRST_IND
# -y
LAST_IND
# -M
TRIM_PROP
# -m
MASK_PROP
# -ex EXCLUDE_LIST
# -rho PROCRUSTES_SCALE
###----End of file----###
5
3.2
Parallel jobs
We currently do not implement multi-thread option to run parallel jobs. Because each study
sample is analyzed independently, users can easily parallel the analyses by running multiple
jobs simultaneously. The -x and -y flags provide a convenient way to specify a subset of
samples to analyze in each job. For example, running the following commands will submit
two jobs: the first job will analyze samples 1 to 350; and the second job will analyze samples
351 to 700 in the HGDP 700 chr22.seq file.
./trace -p ./example/example.conf -x 1 -y 350 -o results.1-350 &
./trace -p ./example/example.conf -x 351 -y 700 -o results.351-700 &
Outputs from these two jobs will have different file name prefixes results.1-350 and results.351700 specified by the -o flag. We also recommend users to provide the coordinatefile when
running multiple jobs using the same set of of reference individuals to save computational time
by avoiding redundant calculation of the reference PCA coordinates in each job.
3.3
Internal reference
In cases when external references are not available or when users want to focus on the study
sample, TRACE provides an option to use a subset of individuals randomly from the study
sample as an internal reference panel. The movitation here is to reduce computational cost
when the sample size is too large for using standard PCA approaches. The size of the internal
reference panel is defined by the parameter REF SIZE (-N). This option is only in effect when
the parameter GENO FILE (-g) is undefined. For example, after commenting out the line
for GENO FILE in the example.conf file, the following command will randomly select 300
individuals from the STUDY FILE as the reference panel, and run the TRACE analysis on
the remaining 400 individuals.
./trace -p ./example/example.conf -N 300
Note that there is trade-off between computational efficiency and accuracy when using this
internal reference option. Setting a smaller value for REF SIZE can reduce the computational
time, but might lead to less accurate results compared to results based on standard PCA. In
addition, the option for running multiple jobs in parallel using parameters FIRST IND (-x)
and LAST IND (-y) is not available when using the internal reference option.
4
Input files
In this section, we describe four input files that are taken by TRACE — the parameterfile,
the studyfile, the genotypefile, and the coordinatefile. The genotypefile and the studyfile have
6
the same format. We also describe one additional file, the sitefile, which is associated with
the the genotypefile and the studyfile. Note that the formats of the genotypefile, the sitefile,
and the coordinatefile are the same as used by the LASER software (Wang et al., 2014).
4.1
parameterfile ( .conf )
The parameterfile contains all parameters required for running TRACE. The default parameterfile is “trace.conf”, which does not need to be explicitly specified in the command line
(i.e. ./trace is equivalent to ./trace -p trace.conf). There are nine parameters in the
parameterfile , including six main parameters and three advanced parameters. Each parameter
is followed by its assigned value, separated by whitespaces. Text in the same line after a ‘#’
character is treated as comment and will not be read. For example, the following parameter
specifications are equivalent in setting the parameter DIM equal to 4:
DIM 4
DIM
4 # Number of PCs to compute
DIM 4
# Other comments
If the user does not assign a value to a parameter in the parameterfile, this parameter
must be followed by a ‘#’ character (even without comments) to avoid unexpected errors in
assigning other parameter values. An example parameterfile is provided in Section 3.1. To
generate an empty template parameterfile, run TRACE when the default parameterfile does
not exist and without any command line arguments. Three parameters do not have default
values, among which two parameters (GENO FILE and STUDY FILE ) need to be explicitly
defined by the users when in use, either in the parameterfile or in the command line (see
Section 5.3), and one parameter (COORD FILE ) is optional. The other six parameters do
not need to be explicitly defined unless the user wants to use settings different from the default.
Please refer to Section 5 for more information on these parameters.
4.2
studyfile ( .geno), genotypefile ( .geno), and sitefile ( .site)
The studyfile and the genotypefile contain genotype data for the study sample and the reference
panel, respectively. TRACE does not require the genotypefile and the studyfile to contain
same set of loci in the same order. Detail information of the loci should be provided in the
sitefile’s. TRACE will automatically extract loci shared by the studyfile and the genotypefile
for downstream analyses.
Each line in the genotypefile/studyfile represents genotype data of one individual. The
first two columns represent population IDs and individual IDs, respectively. Starting from
the third column, each column represent a locus. We only consider bi-allelic SNP markers.
7
To be consistent with the sequence data, genotypes should be given on the forward strand.
Genotypes are coded by 0, 1, or 2, representing copies of the reference allele at a locus in one
diploid individual. Missing data are coded by -9. TRACE can also be applied to multi-ploidy
organisms. In general, genotypes should be coded by 0, 1, ..., K for K-ploidy organisms.
Columns in the genotypefile/studyfile are tab-delimited. An example is provided below:
POP_1
POP_1
POP_2
POP_3
...
IND_1
IND_2
IND_3
IND_4
...
2
2
0
1
...
0
-9
0
2
...
1
2
-9
1
...
...
...
...
...
...
Information on each locus, including chromosome number, genomic position, SNP ID,
reference allele, and alternative allele, is listed in a separate sitefile. The reference allele and
the alternative allele should be given on the forward strand. The first row of the sitefile is
the header line. Starting from the second line, each line represents one locus. Columns in the
sitefile are tab-delimited. An example sitefile is provided below:
CHR
1
1
1
...
POS
752566
768448
1005806
...
ID
rs3094315
rs12562034
rs3934834
...
REF
G
G
C
...
ALT
A
A
T
...
To run TRACE, a “.site” file is required for each “.geno” file. Users can convert their
genotype data from VCF format to our “.geno” and “.site” format using a program called
vcf2geno. This program is available from the LASER software package (http://www.sph.
umich.edu/csg/chaolong/LASER/). The command line for running vcf2geno is
./vcf2geno --inVcf filename.vcf --out output
which will generate a genotypefile named “output.geno” and a sitefile named “output.site”.
4.3
coordinatefile ( .coord )
The coordinatefile contains PCA coordinates of the reference individuals. The first line is
the header line. Starting from the second line, each line represent one individual. The first
two columns correspond to population IDs and individual IDs respectively, and the following
K columns represent the top K principal components (PCs). The order of the reference
individuals must be the same as in the genotypefile. The coordinatefile is required to be
tab-delimited. Below is an illustration of the format of the coordinatefile:
8
popID
POP_1
POP_1
POP_2
POP_3
...
indivID
IND_1
IND_2
IND_3
IND_4
...
PC1
-3.5
-2.2
7.8
1.6
...
PC2
0.2
4.5
-0.8
-3.8
...
PC3
0.7
0.8
-1.0
-0.4
...
...
...
...
...
...
...
If the coordinatefile is not provided, TRACE will automatically compute the reference
coordinates based on the genotypefile, and the results will be output. We recommend users
to prepare a coordinatefile as input for TRACE when submitting multiple jobs using the the
same reference panel (so that the same computation will not be repeated for every job).
5
Usage options
TRACE has 16 parameters that users can set in the parameterfile, including 7 main parameters
that are required for running TRACE and 9 advanced parameters for some special options.
Among the 7 main parameters, 3 are parameters regarding the input data files and need
to be explicitly defined when in use. The other 4 main parameters and the 9 advanced
parameters have default values. In addition, TRACE takes 17 command line arguments,
which are described in Section 5.3.
5.1
Main parameters
STUDY FILE (string) The name of the studyfile. If the file is not in the same directory as
TRACE, the whole path must be specified. This parameter must be explicitly defined.
GENO FILE (string) The name of the genotypefile. If the file is not in the same directory as TRACE, the whole path must be specified. This parameter must be explicitly defined
unless using the internal reference option.
COORD FILE (string) The name of the coordinatefile. If the file is not in the same directory as TRACE, the whole path must be specified. This parameter is optional. If undefined, TRACE will automatically compute the reference coordinates based on the genotypefile.
OUT PREFIX (string) The prefix that will be added to the file names of outputting results. A path can be specified to output results to a different directory. The default value is
“TRACE ”.
9
DIM (int) The number of PCs to compute (must be a positive integer). This number must be
smaller than the number of individuals and the number of loci in the genotypefile, and cannot
be greater than the number of PCs in the coordinatefile if a coordinatefile is provided. The
default value is 2.
DIM HIGH (int) Dimension of the sample-specific PCA map to project from (must be
a positive integer). This number must be smaller than the number of individuals and the
number of loci in the genotypefile, and cannot be smaller than DIM. TRACE will project each
study sample from a DIM HIGH dimensional PC space to the DIM dimensional reference
ancestry map. If set to 0, the program will use the number of significant PCs based on TracyWidom tests for each sample. The default value is 20.
MIN LOCI (int) The minimum number of non-missing loci required for an individual in
the study sample to be analyzed (must be a positive integer). If the number of non-missing
loci is smaller than MIN LOCI, the individual will not be analyzed and results for this individual are output as “NA”. The default value is 100.
5.2
Advanced parameters
ALPHA (double) Significance level in Tracy-Widom tests to determine the number of informative PCs, or DIM HIGH, in the sample-specific PCA (must be a number between 0 and 1).
This parameter is effective only if DIM HIGH is undefined or set to 0. The default value is 0.1.
THRESHOLD (double) Convergence criterion of the projection Procrustes analysis (must
be a positive number). The default value is 0.000001.
REF SIZE (int) The number of individuals to be randomly selected from the study sample as
an internal reference panel (must be a positive integer). This number cannot be greater than
the number of individuals in the studyfile. This parameter will be in effect only if GENO FILE
is undefined (i.e., an external reference panel is not provided). The default value is 200.
FIRST IND (int) The index of the first individual in the study sample to analyze (must
be a positive integer). This number cannot be greater than the number of individuals in
the studyfile. Individuals that have indices smaller than FIRST IND will be skipped. This
parameter will be in effect only if GENO FILE is defined (i.e., an external reference panel is
provided). The default value is 1.
LAST IND (int) The index of the last individual in the study sample to analyze (must
10
be a positive integer). This number cannot be greater than the number of samples in the
studyfile or smaller than FIRST IND. Individuals that have indices greater than LAST IND
will be skipped. This parameter will be in effect only if GENO FILE is defined (i.e., an external reference panel is provided). The default value is the number of samples in the studyfile.
TRIM PROP (double) Proportion of randomly selected loci to exclude from the analysis for all samples (must be a number between 0 and 1). This option is useful when the data
size is too big such that memory limit becomes an issue. The default value is 0.
MASK PROP (double) Proportion of loci in a study sample that will be randomly set
to missing (must be a number between 0 and 1). This option is useful when testing the robustness to missing data or examing the minimal number of markers requested in a sample in
order to have satisfying performance given a reference panel. The default value is 0.
EXCLUDE LIST (string) This parameter specifies the file name of a list of SNPs to exclude from the analysis. Each line of the file is a SNP ID and no header is required. If the file
is not in the same directory as TRACE, the whole path must be specified. This parameter do
not have default value.
PROCRUSTES SCALE (int) This parameter specifies how to scale coordinates in the
Procrustes analysis (must be 0 or 1). When set to 0, the Procrustes analysis will include
the scaling factor as a free parameter and estimates its value to minimize the sum of squared
Euclidean distances between two sets of coordinates in the analysis. When set to 1, the Procrustes analysis will fix the scaling factor so that two sets of coordinates will have the same
total variance (i.e. Procrustes analysis only search for rotation, reflection and translation to
minimize the sum of squared Euclidean distances). The default value is 0.
5.3
Command line arguments
The command line flags provide the user an option to enter information from the command line.
All command line arguments will overwrite values specified in the parameterfile. If a parameter
is specified with an invalid value in the parameterfile but a valid value in the command line,
the program will return a warning message and still execute correctly by taking the value
from the command line. However, if a parameter value in the command line is not valid,
the program will exit with an error message. If a command line flag is specified, it must be
followed by a space and then the parameter value. Different command line flags can appear
in any order. If the same command line flag is defined more than once, only the last value will
be taken. For example, the following command lines are equivalent and will change the value
11
of the parameter DIM, for which the command line flag is -k, to be 4 while using the other
parameters defined in the parameterfile named “my parameterfile”.
./trace -p my_parameterfile -k 4
./trace -k 4 -p my_parameterfile
./trace -k 3 -p my_parameterfile -k 4
Most command line arguments are optional except for the parameterfile, for which the command line flag is -p. A list of all command line flags is provided below.
-p This flag defines the parameterfile. If the parameterfile is not in the current directory,
a whole path to the file must be specified. This parameter can only be defined using the command line. If undefined, the program will use the default parameterfile named “trace.conf”
in the current directory. If this file does not exist, an empty template parameterfile named
“trace.conf” will be created in the current directory, and the program will then exit with an
error message.
-s Change the parameter value of STUDY FILE.
-g Change the parameter value of GENO FILE.
-c Change the parameter value of COORD FILE.
-o Change the parameter value of OUT PREFIX (useful when running parallel jobs).
-k Change the parameter value of DIM.
-K Change the parameter value of DIM HIGH.
-l Change the parameter value of MIN LOCI.
-a Change the parameter value of ALPHA.
-t Change the parameter value of THRESHOLD.
-N Change the parameter value of REF SIZE.
-x Change the parameter value of FIRST IND (useful when running parallel jobs).
12
-y Change the parameter value of LAST IND (useful when running parallel jobs).
-M Change the parameter value of TRIM PROP (useful when memory is limited).
-m Change the parameter value of MASK PROP.
-ex Change the parameter value of EXCLUDE LIST.
-rho Change the parameter value of PROCRUSTES SCALE.
6
Output files
All output files will be saved in the current directory unless the path to a different directory
is given in the parameter value of OUT PREFIX. All output file names will start with the
parameter value of OUT PREFIX. These files are described below.
6.1
.log and terminal outputs
The terminal outputs are used to monitor and record the progress when running TRACE. It
starts with all parameter values used in the execution of TRACE, and reports the progress of
the program step by step. The log file is identical to the terminal outputs.
6.2
.RefPC.coord and .RefPC.var
When COORD FILE is not defined, TRACE will perform PCA on the reference genotype data
given by the genotypefile. Results of the top k PCs, where k is defined by the parameter DIM,
are output to two files named OUT PREFIX.RefPC.coord and OUT PREFIX.RefPC.var.
The RefPC.coord file records the PCA coordinates of the reference individuals. The first
line in this file is a header line. Starting from the second line, each line represents one
individual. The first two columns are population ID and individual ID, respectively. The
remaining columns correspond to the top k PCs. This file is tab-delimited. The format of this
file is exactly the same as the coordinatefile (Section 4.3), so that this file can be directly used
as the input file for TRACE.
The RefPC.var file records the proportion of variance explained by each PC. The first line
in this file is a header line. Starting from the second line, each line represents one PC. The
first column is the PC index and the second column is the percentage of variance explained
by each PC. Only results for the top k PCs are output. This file is tab-delimited.
13
6.3
.ProPC.coord
This file contains the estimated PCA (or “Procrustean PCA”) coordinates of the study sample.
The first line is a header line. Starting from the second line, each line represents one study
sample. The first column is population ID, and the second column is individual ID. The
third column reports the number of nonmissing loci used in the analysis. The fourth column
reports values of DIM HIGH, dimension of the sample-specific PCA map used in projection
Procrustes analysis. The fifth column reports the Procrustes similarity score between each
sample-specific PCA map (after being projected to the DIM dimensional space) and the
original DIM dimensional reference PCA map. Starting from the sixth column, each column
represents coordinates of one PC (up to the kth PC, where k is defined by DIM ). Columns in
this file are tab-delimited.
When using the internal reference option, coordinates for the individuals that are selected
as the internal reference will also be included in this file. The values of DIM HIGH (the fourth
column) and the Procrustes similarity scores (the fifth column) for these individuals will be
listed as “NA”.
7
Computational complexity
TRACE examines one study individual at a time. Therefore, the computational costs increase
linearly with the number of individuals to be analyzed. We can easily run the analyses in
parallel by submitting multiple jobs to analyze different subsets of the study sample (using
command line flags -x and -y).
The cost for analyzing each individual depends on the number of individuals, N , and
the number of loci, L, in the reference panel, and the proportion of missing data, m, in the
study individual. We first calculate the N × N pairwise similarity matrix of the reference
panel, for which the computational cost is O(N 2 L). This computation is only performed once
and will be repeatedly used in analyzing each individual. Roughly, we expect computational
cost of PCA for each study individual to be O(N L + N 2 Lm + N 3 ), in which N L is the
time required for comuputing the extra row (and column) for the study individual in the
similarity matrix, N 2 Lm is for adjustment of the similarity matrix to account for missing
data in the study individual, and N 3 is for eigen decomposition on the similarity matrix. The
computational cost of projection Procrustes analysis is approximately O[q(N K + K 3 )], where
q is the number of iterations required for projection Procrustes analysis to converge and K
is the value of DIM HIGH. When K is not big, the analysis often converges within a few
iterations, and the cost for projection Procrustes analysis is negligible compared to the cost
for PCA. Overall, the computational cost for TRACE on a study sample of n individuals
is approximately O[N 2 L + n(N L + N 2 Lm
¯ + N 3 + qN K + qK 3 )]. If we ignore the cost for
14
projection Procrustes analysis and suppose the missing data rate is small, the computational
cost for TRACE can be approximated as O[N 2 L + n(N L + N 3 )]. In comparison, the cost for
a standard PCA on n individuals is approximately O(n2 L + n3 ).
8
Version changes
Changes from previous versions of the TRACE software are noted here.
8.1
Version 1.0 (May 19, 2014)
- Initial release of the TRACE software.
8.2
Version 1.01 (Apr 6, 2015)
- Added a new parameter PROCRUSTES SCALE to allow scaling two sets of coordinates
to have the same variance in the Procrustes analysis.
- Fixed a minor bug to in the projection Procrustes anlaysis (convergence issue, no impact
on the results).
9
Acknowledgements
I would like to thank Conrad Sanderson at the National ICT Australia for his help on using
the Armadillo library.
References
Gower, J. C., and G. B. Dijksterhuis, 2004
Press.
Procrustes Problems. Oxford University
Lee, A. B., D. Luca, L. Klei, B. Devlin and K. Roeder, 2010
ancestry using spectral graph theory. Genet. Epidemiol. 34: 51–59.
Discovering genetic
Li, J. Z., D. M. Absher, H. Tang, A. M. Southwick, A. M. Casto, S. Ramachandran, H. M. Cann, G. S. Barsh, M. Feldman, L. L. Cavalli-Sforza and R. M.
Myers, 2008 Worldwide human relationships inferred from genome-wide patterns of variation. Science 319: 1100–1104.
McVean, G., 2009 A genealogical interpretation of principal components analysis. PLoS
Genet. 5: e1000686.
15
Patterson, N., A. L. Price and D. Reich, 2006 Population structure and eigenanalysis.
PLoS Genet. 2: 2074–2093.
Sanderson, C., 2010
Armadillo: an open source C++ linear algebra library for fast
protoyping and computationally intensive experiments. Technical Report, NICTA.
¨ m, M. Raghavan, J. Stor˚
Skoglund, P., H. Malmstro
a, P. Hall, E. Willerslev,
¨ therstro
¨ m and M. Jakobsson, 2012 Origins and genetic legacy
M. T. Gilbert, A. Go
of Neolithic farmers and hunter-gatherers in Europe. Science 336: 466–469.
Wang, C., Z. A. Szpiech, J. H. Degnan, M. Jakobsson, T. J. Pemberton, J. A.
Hardy, A. B. Singleton and N. A. Rosenberg, 2010 Comparing spatial maps of
human population-genetic variation using Procrustes analysis. Stat. Appl. Genet. Mol. Biol.
9: Article 13.
Wang, C., X. Zhan, J. Bragg-Gresham, D. Stambolian, E. Chew, K. Branham,
J. Heckenlively, R. S. Fulton, R. K. Wilson, E. R. Mardis, X. Lin, A. Swaroop,
¨ llner and G. R. Abecasis, 2014 Ancestry estimation and control of population
S. Zo
stratification for sequence-based association studies. Nature Genetics 46: 409–415.
Wang, C., X. Zhan, L. Liang, G. R. Abecasis and X. Lin, 2015 Improved ancestry
estimation for both genotyping and sequencing data using projection Procrustes analysis
and genotype imputation. American Journal of Human Genetics (accepted).
16